BERT Based Semi-Supervised Hybrid Approach for Aspect and Sentiment Classification

Aspect-based sentiment analysis (ABSA) includes two sub-tasks namely, aspect extraction and aspect-level sentiment classification. Most existing works address these sub-tasks independently by applying a supervised learning approach using labeled data. However, obtaining such labeled sentences is difficult and extremely expensive. Hence, it is important to solve ABSA without taking a dependency on labeled sentences. In this work, we propose a three-step semi-supervised hybrid approach that jointly detects an aspect and its associated sentiment in a given review sentence. The first step of our approach takes a small set of seed words for each aspect and sentiment class to construct respective semantically coherent class vocabularies. The second step makes use of these constructed vocabularies along with POS tags to label a subset of sentences from the training corpus. As we adopt a semi-automated approach to label the data, this process may induce noise in the labels during the annotation. In the last step, we use such labeled sentences to build a noise-robust deep neural network for aspect and sentiment classification. We conduct experiments on two real data sets to verify the effectiveness of our model (https://github.com/Raghu150999/UnsupervisedABSA).